Fast neural "radio map" models take a top-down map of buildings plus a transmitter location and predict, in about ten milliseconds, how much signal is lost at every street-level point. Running the physics directly with ray tracing takes around a hundred seconds per scene. The weakness of these models is that they break when moved to a city they were not trained on. In the ICASSP 2023 challenge, every top model trained on six European and Mediterranean cities lost substantial accuracy on Istanbul. We argue that part of this failure comes from the coordinate frame the network sees. Radio propagation is organized along rays leaving the transmitter: shadows begin at building edges and extend radially, and loss accumulates with each wall a ray crosses. A standard convolutional network on a north-up grid has to rediscover this ray structure from street layouts. Street layouts vary by city (rectangular grids versus organic, irregular blocks) much more than ray geometry does.
The proposal is to resample the scene into polar coordinates centred on the transmitter, so each row is a ray and each column a distance. In that frame, occlusion, line-of-sight and the distance to the blocking edge can be computed exactly with a cumulative maximum along each row, in well under a millisecond. Rotating the city becomes a circular shift. A network in this frame learns relations along rays that do not depend on whether streets run north-south. Polar resampling blurs detail far from the transmitter, so we pair the polar trunk with a thin high-resolution Cartesian path that carries building edges. The two views are fused before the output. The claim is testable and could be false: if a Cartesian network given the same physics channels and rotation augmentation closes the cross-city gap equally well, the polar frame adds nothing.
Core Claim
Holding inputs, training data and latency budget fixed, a transmitter-centred polar surrogate with a Cartesian detail path will shrink the error gap between training cities and an unseen city. The improvement should concentrate in a narrow band around shadow boundaries, which is where diffraction makes signal strength change abruptly. The mechanism has two parts. First, ray-aligned computation lets exact line-of-sight and edge-distance features be computed per ray, and lets convolutions run along the direction physics actually acts. Second, a rotation of the city about the transmitter becomes a circular shift in angle, and circular convolutions handle shifts exactly. The model should stay inside the roughly 10 ms window that RadioUNet and the challenge winners established.
For a rooftop transmitter at height $h_{tx}$ and a receiver at height $h_{rx}$ at polar cell $(\theta, r)$, line-of-sight holds exactly when
$$ \frac{h_{rx}-h_{tx}}{r} \;\ge\; \max_{r'<r} \frac{h(\theta,r')-h_{tx}}{r'}, $$
where $h(\theta,r')$ is the building height along the ray at angle $\theta$ and distance $r'$. The right-hand side is a running maximum along each polar row. The distance $r'$ at which that maximum is reached gives the position of the dominant obstructing edge, and therefore a knife-edge diffraction geometry feature at no extra cost. A running sum of building cells gives a wall-crossing count. All three features are one-pass scans on a GPU.
Relation to Prior Work
Three lines of work motivate the gap. RadioUNet introduced the geometry-to-pathloss U-Net with roughly 1 dB error at $10^{-3}$ to $10^{-2}$ s, but evaluated inside one dataset distribution. RMTransformer added multi-axis attention for long-range interactions and skip connections for alley-scale detail, reaching RMSE 0.007148 on the USC dataset, but tested only that single campus footprint. The challenge papers (The First Pathloss Radio Map Prediction Challenge, Overview) documented the Istanbul generalization gap without resolving it. All three work in a north-up Cartesian frame. PMNet's dilated convolutions are in effect a way to widen the receptive field enough to capture ray-scale effects. We change the frame instead of widening the receptive field.
Two results constrain the design. The challenge's second-place entry, Agile, already fed fractional line-of-sight maps as input and still trailed PMNet (RMSE 0.0451 versus 0.0383). So simply adding a physics channel is not enough, and our decisive ablation must give the Cartesian baselines exactly the same channels. Separately, STAR handles spatial heterogeneity through repeated spatial scanning and network branching. RadioLAM handles heterogeneity by sampling 64 diffusion candidates and choosing among them with a physics check. Both are too slow for millisecond inference, which rules out adapting either at test time.
The closest work on invariance is FM-RME. It builds rotation and translation symmetry into a graph-network feature extractor and shows zero-shot transfer to an unseen dataset. It reconstructs maps from sparse measurements, however, and is not aimed at sharp building-edge transitions. Our model maps geometry directly to a dense map, gets symmetry only about the transmitter, which is the one centre that matters physically, and targets boundary fidelity explicitly. PI-DRME separates shared log-distance physics from local shadowing, but its reconstructions are smoothed at edges. The polar trunk plays the role of its shared physics component, while the Cartesian path preserves the edges it loses.
From remote sensing, Accurate Building Detection in VHR Remote Sensing Images using Geometric Saliency found that mid-level geometric cues such as corners and junctions transfer across regions where deep backbones collapse (F-score falling from 0.74 to 0.10). BASeg showed that isotropic pixel losses blur edges across ten global cities. We take two things from these: the hypothesis that geometry, not appearance, should carry transfer, and the decision to measure error in a boundary band rather than only on average.
Research Program
The scene representation stacks building height, a building mask and transmitter distance. It is warped to a polar grid of 1024 angles by 256 log-spaced radii using a differentiable bilinear sampler. At the 256-pixel grid corner, an angular bin spans about one pixel, so far-field sampling stays close to native resolution. Three channels are then computed per ray: exact line-of-sight from the running maximum above, the distance to the dominant obstructing edge, and the cumulative wall count.
The polar trunk is a small encoder-decoder. It uses circular padding in angle and anisotropic kernels that are longer along range than across angle. Its output features are warped back to Cartesian coordinates at two scales. The Cartesian path is a shallow, full-resolution convolutional branch on the raw building map. Its job is the street-building interface and narrow gaps. The two feature sets are concatenated and passed to a light output head. Training uses the challenge's masked mean-squared error.
Beyond the cross-city comparison, we will analyse where the error lives. We will build a per-pixel error decomposition by distance to the nearest line-of-sight/non-line-of-sight transition, using the exact line-of-sight map as a reference. This tests the premise that the cross-city gap concentrates at shadow boundaries; if it does not, the boundary-band argument fails. We will also apply controlled perturbations: rotating Istanbul scenes, and synthetically regularizing or irregularizing street orientation. These isolate whether orientation statistics drive the gap.
Baselines
The naive reference is a log-distance pathloss fit plus the exact line-of-sight mask. It is a two-regime empirical model with no learning beyond two slopes, and it shows how much of the problem cheap physics already solves. Next is RadioUNet, the standard Cartesian U-Net, retrained on the identical six-city split.
The strongest competitor on this exact benchmark is PMNet, the challenge winner. It reports RMSE 0.0383 on the Istanbul evaluation set (The First Pathloss Radio Map Prediction Challenge). It is the right opponent because it already attacks long-range propagation through dilation and meets the latency budget. Beating it means lower Istanbul RMSE than our faithful reimplementation of PMNet trained under our pipeline, and, where an evaluation route to the official Istanbul labels exists, lower RMSE than the reported 0.0383. If the hidden labels are not accessible, we compare only against our retrained PMNet and say so.
RMTransformer is the strongest in-distribution model (RMSE 0.007148 on USC). We retrain it on the six cities to test whether attention capacity alone transfers. For fairness, every model gets identical inputs including the three ray features, which we warp back to Cartesian for Cartesian models. Every model also gets the same rotation and flip augmentation, the same epochs, and the same hyperparameter search of twenty trials on a held-out training city. Parameter counts are matched within 20%, and all models must run under 10 ms at batch size one on the same GPU.
The key ablations remove each ingredient in turn. The first is the polar trunk alone, without the Cartesian detail path, which isolates detail preservation. The second is the full model with zero-padded rather than circular angular convolutions, which isolates the symmetry. The third is the full model with the ray features removed, which isolates exact physics features from the frame itself.
Why not just…?
Why not just add line-of-sight channels to PMNet? This is the main threat, and it is close to what Agile did. Agile's fractional line-of-sight input still left it at 0.0451 against PMNet's 0.0383. That suggests the channel alone is not decisive, but Agile's channel was an approximation, while ours is exact and comes with edge-distance and wall-count features. The decisive experiment gives PMNet and RadioUNet our exact channels. If either then matches the polar model on Istanbul and in the boundary band, we will report that the features, not the frame, carry the gain, and we will narrow the claim accordingly.
Why not just use heavy rotation augmentation on a Cartesian network? Augmentation can teach approximate rotation invariance. It cannot make a Cartesian convolution compute along rays, and ray-wise accumulation is what dilated convolutions in PMNet approximate. Every baseline receives the same augmentation, so this alternative is built into the comparison. The rotated-Istanbul test will also show whether the augmented baselines are actually invariant: their error should be flat across rotation angles, whereas a polar model with circular convolutions is exactly invariant to rotation about the transmitter, apart from interpolation.
Why not just train on more geographically diverse cities? Trusting Small Training Dataset for Supervised Change Detection found that diversity beat proximity (Kappa 0.32 versus 0.12), so this is plausible and complementary rather than competing. We test it directly. SpectrumNet's dense-urban, urban, suburban and rural splits let us vary training diversity for every model and check whether the polar advantage persists or vanishes as diversity grows. If the gap vanishes, the contribution reduces to data efficiency, and we would say so.
Why not just fine-tune on a few ray-traced maps from the target city? That requires running a ray tracer in every new city, which is the cost the surrogate exists to avoid. We report it as an upper reference: few-shot fine-tuning with 10 Istanbul maps, against which zero-shot transfer is measured.
Proposed Experiments
The primary experiment uses RadioMap3DSeer as configured for the challenge. Models train on 701 maps from Ankara, Berlin, Glasgow, Ljubljana, London and Tel Aviv. We test on Istanbul where labels are available, and otherwise run leave-one-city-out across the six training cities, holding out Tel Aviv and Ankara as the most morphologically distinct.
The metrics are challenge-normalized RMSE, RMSE in bands within 2, 5 and 10 pixels of the exact shadow boundary, the ratio of held-out-city RMSE to in-distribution RMSE, and wall-clock latency at batch size one. Secondary checks have two goals. The first is to confirm the polar model gives up nothing in distribution, using the USC split where RMTransformer's 0.007148 is the reference. The second is to test transfer across environment types on SpectrumNet at 1.5 m height; RadioLAM reports only relative MSE reductions there, so no absolute reference number was found.
As a downstream check, we feed predicted maps for held-out cities into LocUNet and measure localization error. LocUNet reports 4.73 m MAE with accurate maps and 13.15 m under worst-case map mismatch, so we can see whether cross-city map errors translate into positioning errors.
The claim is supported only if three conditions hold together. The polar model must beat the feature- and augmentation-matched PMNet on held-out-city RMSE. Most of that gain must sit in the boundary bands. Removing either circular padding or the Cartesian path must erase a measurable share of the gain. A win on average RMSE alone, without the boundary-band result, would not support the mechanism we propose.
References
- RadioUNet: Fast Radio Map Estimation with Convolutional Neural Networks
- RMTransformer: Accurate Radio Map Construction and Coverage Prediction
- The First Pathloss Radio Map Prediction Challenge
- Overview of the First Pathloss Radio Map Prediction Challenge
- Statistically-Guided Deep Network Transformation to Harness Heterogeneity in Space (Extended Abstract)
- RadioLAM: A Large AI Model for Fine-Grained 3D Radio Map Estimation
- FM-RME: Foundation Model Empowered Radio Map Estimation
- Physics-Inspired Distributed Radio Map Estimation
- Accurate Building Detection in VHR Remote Sensing Images Using Geometric Saliency
- BASeg: Boundary-Aware Remote Sensing Segmentation with Structural Penalties
- Trusting small training dataset for supervised change detection
- LocUNet: Fast Urban Positioning Using Radio Maps and Deep Learning